Search

The research groups in computer vision, graphics, and machine learning have dedicated a substantial amount of attention to the areas of 3D object reconstruction, augmentation, and registration. Deep learning is the predominant method used in artificial intelligence for addressing computer vision challenges. However, deep learning on three-dimensional data presents distinct obstacles and is now in its nascent phase. There have been significant advancements in deep learning specifically for three-dimensional data, offering a range of ways to address these issues. This study offers a comprehensive examination of the latest advancements in deep learning methodologies. We examine many benchmark models for the tasks of 3D object registration, augmentation, and reconstruction. We thoroughly analyse their architectures, advantages, and constraints. In summary, this report provides a comprehensive overview of recent advancements in three-dimensional deep learning and highlights unresolved research areas that will need to be addressed in the future. [ABSTRACT FROM AUTHOR]

Machine learning (ML), a form of artificial intelligence (AI), is being increasingly employed in neurology. Reported performance metrics often match or exceed the efficiency of average clinicians. The neurologist is easily baffled by the underlying concepts and terminologies associated with ML studies. The superlative performance metrics of ML algorithms often hide the opaque nature of its inner workings. Questions regarding ML model's interpretability and reproducibility of its results in real-world scenarios, need emphasis. Given an abundance of time and information, the expert clinician should be able to deliver comparable predictions to ML models, a useful benchmark while evaluating its performance. Predictive performance metrics of ML models should not be confused with causal inference between its input and output. ML and clinical gestalt should compete in a randomized controlled trial before they can complement each other for screening, triaging, providing second opinions and modifying treatment. [ABSTRACT FROM AUTHOR]

This paper presents a new edge detection process implemented in an embedded IoT device called Bee Smart Detection node to detect catastrophic apiary events. Such events include swarming, queen loss, and the detection of Colony Collapse Disorder (CCD) conditions. Two deep learning sub-processes are used for this purpose. The first uses a fuzzy multi-layered neural network of variable depths called fuzzy-stranded-NN to detect CCD conditions based on temperature and humidity measurements inside the beehive. The second utilizes a deep learning CNN model to detect swarming and queen loss cases based on sound recordings. The proposed processes have been implemented into autonomous Bee Smart Detection IoT devices that transmit their measurements and the detection results to the cloud over Wi-Fi. The BeeSD devices have been tested for easy-to-use functionality, autonomous operation, deep learning model inference accuracy, and inference execution speeds. The author presents the experimental results of the fuzzy-stranded-NN model for detecting critical conditions and deep learning CNN models for detecting swarming and queen loss. From the presented experimental results, the stranded-NN achieved accuracy results up to 95%, while the ResNet-50 model presented accuracy results up to 99% for detecting swarming or queen loss events. The ResNet-18 model is also the fastest inference speed replacement of the ResNet-50 model, achieving up to 93% accuracy results. Finally, cross-comparison of the deep learning models with machine learning ones shows that deep learning models can provide at least 3–5% better accuracy results. [ABSTRACT FROM AUTHOR]

Context: With the union of machine learning (ML) and quantum chemistry, amid the debate between machine-learned functionals and human-designed functionals in density functional theory (DFT), this paper aims to demonstrate the generation of potential energy surfaces using computations with machine-learned density functional approximation (ML-DFA). A recent research trend is the application of ML in quantum sciences in the design of density functionals such as DeepMind's Deep Learning model (DeepMind21, DM21). Though science reported the state-of-the-art performance of DM21, the opportunity to utilize DeepMind's pretrained DM21 neural networks in computations in quantum chemistry has not yet been tapped. So far in the literature, the Deep Learning density functionals (DM21) have not been applied to generate potential energy surfaces. While the superior accuracy of DM21 has been reported, there is still a scarcity of publications that apply DM21 in calculations in the field. In this context, for the first time in literature, neural density functionals inferring 2D potential energy surfaces (ML-DFA-PES) based on machine-learned DFA-based computational method is contributed in this paper. This paper reports the ML-DFA-generated PES for C4H8, H2O, H2, and H2+ by employing a pretrained DM21m TensorFlow model with cc-pVDZ basis set. In addition, we also analyze the long-range behavior of DM21 based PES to investigate the ability to describe a system at long ranges. Furthermore, we compare PES diagrams from DM21 with popular DFT functionals (b3lyp/ PW6B95) and CCSD(T). Methods: In this method, 2D potential energy surfaces are obtained using a method that relies upon the neural network's ability to accurately learn the mapping between 3D electron density and exchange–correlation potential. By inserting Deep Learning inference in DFT with a pretrained neural network, self-consistent field (SCF) energy at different geometries along the coordinates of interest is computed, and then, potential energy surfaces are plotted. In this method, first, the electron density is computed mathematically, and this computed 3D electron density is used as a ML feature vector to predict the exchange correlation potential as a ML inference computed by a forward pass of pre-trained DM21 TensorFlow computational graph, followed by the computation of self-consistent field energy at multiple geometries, and then, SCF energies at different bond lengths/angles are plotted as 2D PES. We implement this in a python source code using frameworks such as PySCF and DM21. This paper contributes this implementation in open source. The source code and DM21-DFA-based PES are contributed at https://sites.google.com/view/MLfunctionals-DeepMind-PES. [ABSTRACT FROM AUTHOR]

Recently, transfer learning has gained popularity in the machine learning community. Transfer Learning (TL) has emerged as a promising paradigm that leverages knowledge learned from one or more related domains to improve prediction accuracy in a target domain with limited data. However, for time series forecasting (TSF) applications, transfer learning is relatively new. This paper addresses the need for empirical studies as identified in recent reviews advocating the need for practical guidelines for Transfer Learning approaches and method designs for time series forecasting. The main contribution of this paper is the suggestion of a comprehensive framework for Transfer Learning Sensitivity Analysis (SA) for time series forecasting. We achieve this by identifying various parameters seen from various angles of transfer learning applied to time series, aiming to uncover factors and insights that influence the performance of transfer learning in time series forecasting. Undoubtedly, symmetry appears to be a core aspect in the consideration of these factors and insights. A further contribution is the introduction of four TL performance metrics encompassed in our framework. These TL performance metrics provide insight into the extent of the transferability between the source and the target domains. Analyzing whether the benefits of transferred knowledge are equally or unequally accessible and applicable across different domains or tasks speaks to the requirement of symmetry or asymmetry in transfer learning. Moreover, these TL performance metrics inform on the possibility of the occurrence of negative transfers and also provide insight into the possible vulnerability of the network to catastrophic forgetting. Finally, we discuss a sensitivity analysis of an Ensemble TL technique use case (with Multilayer Perceptron models) as a proof of concept to validate the suggested framework. While the results from the experiments offer empirical insights into various parameters that impact the transfer learning gain, they also raise the question of network dimensioning requirements when designing, specifically, a neural network for transfer learning. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNN) have fundamentally revolutionized the artificial intelligence (AI) field. The transformer model is a type of DNN that was originally used for the natural language processing tasks and has since gained more and more attention for processing various kinds of sequential data, including biological sequences and structured electronic health records. Along with this development, transformer-based models such as BioBERT, MedBERT, and MassGenie have been trained and deployed by researchers to answer various scientific questions originating in the biomedical domain. In this paper, we review the development and application of transformer models for analyzing various biomedical-related datasets such as biomedical textual data, protein sequences, medical structured-longitudinal data, and biomedical images as well as graphs. Also, we look at explainable AI strategies that help to comprehend the predictions of transformer-based models. Finally, we discuss the limitations and challenges of current models, and point out emerging novel research directions. [ABSTRACT FROM AUTHOR]

CFRP hybrid bonded–bolted (HBB) joints combine the advantages of traditional joining methods, namely adhesive bonding, and bolting, to achieve optimal connection performance, making them the most favored connection method. The structural parameters of CFRP HBB joints, including overlap length, bolt-hole spacing, and fit clearance relationships, have a complex impact on connection performance. To enhance the connectivity performance of joint structures, this paper develops a multiscale finite element analysis model to investigate the impact of structural parameters on the strength of CFRP HBB joint structures. Coupled with experimental validation, the study reveals how changes in structural parameters affect the unidirectional tensile failure force of the joints. Building on this, an analytical approach and inverse design methodology for the mechanical properties of CFRP HBB joints based on deep supervised learning algorithms are developed. Neural networks accurately and efficiently predict the performance of joints with unprecedented combinations of parameters, thus expediting the inverse design process. This research combines experimentation and multiscale finite element analysis to explore the unknown relationships between the mechanical properties of CFRP HBB joints and their structural parameters. Furthermore, leveraging DNN neural networks, a rapid calculation method for the mechanical properties of hybrid joints is proposed. The findings lay the groundwork for the broader application and more intricate design of composite materials and their connection structures. [ABSTRACT FROM AUTHOR]

There have been a number of approaches to employ neural networks in self-adaptive systems; in many cases, generic neural networks and deep learning are utilized for this purpose. When this approach is to be applied to improve an adaptation process initially driven by logical adaptation rules, the problem is that (1) these rules represent a significant and tested body of domain knowledge, which may be lost if they are replaced by a neural network, and (2) the learning process is inherently demanding given the black-box nature and the number of weights in generic neural networks to be trained. In this paper, we introduce the rule-specific neural network method that makes it possible to transform the guard of an adaptation rule into a rule-specific neural network, the composition of which is driven by the structure of the logical predicates in the guard. Our experiments confirmed that the black box effect is eliminated, the number of weights is significantly reduced, and much faster learning is achieved whilst the accuracy is preserved. This text is an extended version of the paper presented at the ISOLA 2022 conference (Bureš et al. in Proceedings of ISOLA 2022, Rhodes, Greece, pp. 215–230, 2022). [ABSTRACT FROM AUTHOR]

Tsunami simulations require powerful computational resources to be performed efficiently. Although the modern graphics processing units (GPUs) allow the acceleration of this kind of simulations, they can still last many minutes or even hours for simulations which have to deal with very high spatial resolutions or simulation times. In this paper, we propose a method to predict the alert or inundation level of a tsunami generated by an earthquake using deep learning methods. In particular, we train multilayer perceptron (MLP) neural networks for predicting the alert level due to a tsunami at given coastal locations. Ensemble methods are used to improve the predictions of the neural networks. Tsunamis caused by ruptures of several fault segments at different time instants, application to real events, probabilistic forecasting and comparison with other machine learning algorithms are also addressed. Results on realistic scenarios confirm that good accuracies are obtained. The inference times of the trained networks and ensembles are also very low, lasting less than one second to predict the results of thousands of simulations. The proposed method could be used in a tsunami early warning system along with the application of scaling laws. Plain Language Summary: In tsunami early warning systems (TEWS) it is crucial to perform fast predictions of the risk of a tsunami in coastal areas in order to take appropriate measures before the arrival of the tsunami. Currently, tsunami simulations may take many minutes or even hours despite using modern computational resources. In this work we propose a method based on deep learning techniques to predict the alert level of a tsunami generated by an earthquake at given coastal locations in a fraction of a second. This method could be used in TEWS along with techniques to estimate the values of some parameters of the earthquake. Application to real events and probabilistic predictions are also addressed. Key Points: A deep learning method to predict different tsunami alert levels is proposedTsunamis caused by the rupture of multiple fault segments at different time instants are supportedApplication to real events with probabilistic predictions are addressed [ABSTRACT FROM AUTHOR]

For grid stability, operation, and planning, solar irradiance forecasting is crucial. In this paper, we provide a method for predicting the Global Horizontal Irradiance (GHI) mean values one hour in advance. Sky images are utilized for training the various forecasting models along with measured meteorological data in order to account for the short-term variability of solar irradiance, which is mostly caused by the presence of clouds in the sky. Additionally, deep learning models like the multilayer perceptron (MLP), convolutional neural networks (CNN), long short-term memory (LSTM), or their hybridized forms are widely used for deterministic solar irradiance forecasting. The implementation of probabilistic solar irradiance forecasting, which is gaining prominence in grid management since it offers information on the likelihood of different outcomes, is another task we carry out using quantile regression. The novelty of this paper lies in the combination of a hybrid deep learning model (CNN-LSTM) with quantile regression for the computation of prediction intervals at different confidence levels. The training of the different machine learning algorithms is performed over a year's worth of sky images and meteorological data from the years 2019 to 2020. The data were measured at the University of French Polynesia (17.5770° S, 149.6092° W), on the island of Tahiti, which has a tropical climate. Overall, the hybrid model (CNN-LSTM) is the best performing and most accurate in terms of deterministic and probabilistic metrics. In addition, it was found that the CNN, LSTM, and ANN show good results against persistence. [ABSTRACT FROM AUTHOR]

Deep learning is a powerful machine learning technique that can learn from large amounts of data using multiple layers of artificial neural networks. This paper reviews some applications of deep learning in bioinformatics, a field that deals with analyzing and interpreting biological data. We first introduce the basic concepts of deep learning and then survey the recent advances and challenges of applying deep learning to various bioinformatics problems, such as genome sequencing, gene expression analysis, protein structure prediction, drug discovery, and disease diagnosis. We also discuss future directions and opportunities for deep learning in bioinformatics. We aim to provide an overview of deep learning so that bioinformaticians applying deep learning models can consider all critical technical and ethical aspects. Thus, our target audience is biomedical informatics researchers who use deep learning models for inference. This review will inspire more bioinformatics researchers to adopt deep-learning methods for their research questions while considering fairness, potential biases, explainability, and accountability. [ABSTRACT FROM AUTHOR]

Machine learning (ML) and deep learning (DL), in particular, are common tools for anomaly detection (AD). With the rapid increase in the number of Internet-connected devices, the growing desire for Internet of Things (IoT) devices in the home, on our person, and in our vehicles, and the transition to smart infrastructure and the Industrial IoT (IIoT), anomaly detection in these devices is critical. This paper is a survey of anomaly detection in sensor networks/the IoT. This paper defines what an anomaly is and surveys multiple sources based on those definitions. The goal of this survey was to highlight how anomaly detection is being performed on the Internet of Things and sensor networks, identify anomaly detection approaches, and outlines gaps in the research in this domain. [ABSTRACT FROM AUTHOR]

Condition assessment for critical infrastructure is a key factor for the wellbeing of the modern human. Especially for the electricity network, specific components such as oil-immersed power transformers need to be monitored for their operating condition. Classic approaches for the condition assessment of oil-immersed power transformers have been proposed in the past, such as the dissolved gases analysis and their respective concentration measurements for insulating oils. However, these approaches cannot always correctly (and in many cases not at all) classify the problems in power transformers. In the last two decades, novel approaches are implemented so as to address this problem, including artificial intelligence with neural networks being one form of algorithm. This paper focuses on the implementation of an adaptive number of layers and neural networks, aiming to increase the accuracy of the operating condition of oil-immersed power transformers. This paper also compares the use of various activation functions and different transfer functions other than the neural network implemented. The comparison incorporates the accuracy and total structure size of the neural network. [ABSTRACT FROM AUTHOR]

Simple Summary: As COVID-19 shifts from pandemic to endemic, emerging variants may be more or less virulent. Predicting whether an emerging COVID-19 variant has of high risk of causing severe disease is needed to plan for potential burdens on hospital capacity and protecting vulnerable populations. However, it takes time to do laboratory and animal experiments to determine whether a new genetic variant might be more severe, and the results may not be representative of when the virus infects humans. By the time there is epidemiological data on the severity of disease associated with a new variant, it can be too late for designing an optimal public health response. There is a critical need for computer models that can predict severe disease risk from genetic sequence data, which can be obtained from just the first few infections in a potential incoming wave. Two key challenges make computer modeling difficult: (1) sequence changes are complex, and (2) using historical data to predict future disease requires accounting for the confounding effects of changing patient demographics, improving therapeutics, and increased vaccination. In this paper, we introduce a novel interpretable deep learning architecture to solve this problem, demonstrating that it can make robust predictions for emerging variants. Through the COVID-19 pandemic, SARS-CoV-2 has gained and lost multiple mutations in novel or unexpected combinations. Predicting how complex mutations affect COVID-19 disease severity is critical in planning public health responses as the virus continues to evolve. This paper presents a novel computational framework to complement conventional lineage classification and applies it to predict the severe disease potential of viral genetic variation. The transformer-based neural network model architecture has additional layers that provide sample embeddings and sequence-wide attention for interpretation and visualization. First, training a model to predict SARS-CoV-2 taxonomy validates the architecture's interpretability. Second, an interpretable predictive model of disease severity is trained on spike protein sequence and patient metadata from GISAID. Confounding effects of changing patient demographics, increasing vaccination rates, and improving treatment over time are addressed by including demographics and case date as independent input to the neural network model. The resulting model can be interpreted to identify potentially significant virus mutations and proves to be a robust predctive tool. Although trained on sequence data obtained entirely before the availability of empirical data for Omicron, the model can predict the Omicron's reduced risk of severe disease, in accord with epidemiological and experimental data. [ABSTRACT FROM AUTHOR]

Currently, the task of maintaining cybersecurity and reliability in various computer systems is relevant. This problem can be solved by detecting anomalies in the log data, which are represented as a stream of textual descriptions of events taking place. For these purposes, reduction to a One-class classification problem is used. Standard One-class classification methods do not achieve good results. Deep learning approaches are more effective. However, they are not robust to outliers and require a lot of computational effort. In this paper, we propose a new robust approach based on a convolutional autoencoder using fuzzy clustering. The proposed approach uses a parallel convolution operation to feature extraction, which makes it more efficient than the currently popular Transformer architecture. In the course of the experiments, the proposed approach showed the best results for both the cybersecurity and the reliability problems compared to existing approaches. It was also shown that the proposed approach is robust to outliers in the training set. [ABSTRACT FROM AUTHOR]

Breast cancer is among the most common cancers found in women, causing cancer-related deaths and making it a severe public health issue. Early prediction of breast cancer can increase the chances of survival and promote early medical treatment. Moreover, the accurate classification of benign cases can prevent cancer patients from undergoing unnecessary treatments. Therefore, the accurate and early diagnosis of breast cancer and the classification into benign or malignant classes are much-needed research topics. This paper presents an effective feature engineering method to extract and modify features from data and the effects on different classifiers using the Wisconsin Breast Cancer Diagnosis Dataset. We then use the feature to compare six popular machine-learning models for classification. The models compared were Logistic Regression, Random Forest, Decision Tree, K-Neighbors, Multi-Layer Perception (MLP), and XGBoost. The results showed that the Decision Tree model, when applied to the proposed feature engineering, was the best performing, achieving an average accuracy of 98.64%. [ABSTRACT FROM AUTHOR]

Some fundamental visual features have been found to be fully extracted before reaching the cerebral cortex. We focus on direction-selective ganglion cells (DSGCs), which exist at the terminal end of the retinal pathway, at the forefront of the visual system. By utilizing a layered pathway composed of various relevant cells in the early stage of the retina, DSGCs can extract multiple motion directions occurring in the visual field. However, despite a considerable amount of comprehensive research (from cells to structures), a definitive conclusion explaining the specific details of the underlying mechanisms has not been reached. In this paper, leveraging some important conclusions from neuroscience research, we propose a complete quantified model for the retinal motion direction selection pathway and elucidate the global motion direction information acquisition mechanism from DSGCs to the cortex using a simple spiking neural mechanism. This mechanism is referred to as the artificial visual system (AVS). We conduct extensive testing, including one million sets of two-dimensional eight-directional binary object motion instances with 10 different object sizes and random object shapes. We also evaluate AVS's noise resistance and generalization performance by introducing random static and dynamic noises. Furthermore, to thoroughly validate AVS's efficiency, we compare its performance with two state-of-the-art deep learning algorithms (LeNet-5 and EfficientNetB0) in all tests. The experimental results demonstrate that due to its highly biomimetic design and characteristics, AVS exhibits outstanding performance in motion direction detection. Additionally, AVS possesses biomimetic computing advantages in terms of hardware implementation, learning difficulty, and parameter quantity. [ABSTRACT FROM AUTHOR]

This paper concerns the development of a machine learning tool to detect anomalies in the molecular structure of Gallium Arsenide. We employ a combination of a CNN and a PCA reconstruction to create the model, using real images taken with an electron microscope in training and testing. The methodology developed allows for the creation of a defect detection model, without any labeled images of defects being required for training. The model performed well on all tests under the established assumptions, allowing for reliable anomaly detection. To the best of our knowledge, such methods are not currently available in the open literature; thus, this work fills a gap in current capabilities. [ABSTRACT FROM AUTHOR]

Despite continuous improvements, precipitation forecasts are still not as accurate and reliable as those of other meteorological variables. A major contributing factor to this is that several key processes affecting precipitation distribution and intensity occur below the resolved scale of global weather models. Generative adversarial networks (GANs) have been demonstrated by the computer vision community to be successful at super‐resolution problems, that is, learning to add fine‐scale structure to coarse images. Leinonen et al. (2020, https://doi.org/10.1109/TGRS.2020.3032790) previously applied a GAN to produce ensembles of reconstructed high‐resolution atmospheric fields, given coarsened input data. In this paper, we demonstrate this approach can be extended to the more challenging problem of increasing the accuracy and resolution of comparatively low‐resolution input from a weather forecasting model, using high‐resolution radar measurements as a "ground truth." The neural network must learn to add resolution and structure whilst accounting for non‐negligible forecast error. We show that GANs and VAE‐GANs can match the statistical properties of state‐of‐the‐art pointwise post‐processing methods whilst creating high‐resolution, spatially coherent precipitation maps. Our model compares favorably to the best existing downscaling methods in both pixel‐wise and pooled CRPS scores, power spectrum information and rank histograms (used to assess calibration). We test our models and show that they perform in a range of scenarios, including heavy rainfall. Plain Language Summary: The processes that lead to precipitation (rainfall) happen on a very small scale. Weather forecast computer models work on much larger scales, so rainfall is often poorly predicted. In this paper, we develop a method that enhances the resolution of rainfall forecasts by a factor of 10, and makes the forecasts more accurate. We generate many samples of what the rainfall pattern could be, which gives an idea of the uncertainty in the forecast. Our method is based on machine learning and neural networks, which means that we use many past examples of weather forecasts, together with the rainfall that actually happened, and our method "automatically" learns how the forecasts can be improved. We use an existing idea called "Generative Adversarial Networks," which has been used very successfully in image‐related tasks, such as producing realistic higher‐resolution images from low‐resolution ones. Our task is similar to producing a high‐resolution image from a low‐resolution one, hence this approach is promising. Our method outperforms a variety of existing approaches, and even produces good predictions for the most extreme rainfall situations in our data set. These are the scenarios that cause the most real‐world disruption, the most useful events to produce good forecasts for. Key Points: We use generative adversarial neural networks to post‐process global weather forecast model output over the UKWe produce more realistic precipitation forecasts than the input forecast data, at 10X resolution, with excellent statistical propertiesWe match or outperform a state‐of‐the‐art pointwise downscaling scheme, while also producing spatially coherent images [ABSTRACT FROM AUTHOR]

This paper presents a summary and meta-analysis of the first three iterations of the annual International Verification of Neural Networks Competition (VNN-COMP), held in 2020, 2021, and 2022. In the VNN-COMP, participants submit software tools that analyze whether given neural networks satisfy specifications describing their input-output behavior. These neural networks and specifications cover a variety of problem classes and tasks, corresponding to safety and robustness properties in image classification, neural control, reinforcement learning, and autonomous systems. We summarize the key processes, rules, and results, present trends observed over the last three years, and provide an outlook into possible future developments. [ABSTRACT FROM AUTHOR]

Modern smart grids are built based on top of advanced computing and networking technologies, where condition monitoring relies on secure cyberphysical connectivity. Over the network infrastructure, transported data containing confidential information, must be protected as smart grids are vulnerable and subject to various cyberattacks. Various machine learning based classifiers were proposed for intrusion detection in smart grids. However, each of them has respective advantage and disadvantages. Aiming to improve the performance of existing machine learning based classifiers, this paper proposes an adaptive deep learning algorithm with a data pre-processing module, a neural network pre-training module and a classifier module, which work together classify intrusion data types using their high-dimensional data features. The proposed Adaptive Deep Learning (ADL) algorithm obtains the number of layers and the number of neurons per layer by determining the characteristic dimension of the network traffic. With transfer learning, the proposed ADL algorithm can extract the original data dimensions and obtain new abstract features. By combining deep learning models with traditional machine learning-based classification models, the performance of classification of network traffic data is significantly improved. By using the Network Security Laboratory-Knowledge Discovery in Databases (NSL-KDD) dataset, experimental results show that the proposed ADL algorithm improves the effectiveness of existing intrusion detection methods and reduces the training time, indicating a promising candidate to enhance network security in smart grids. [ABSTRACT FROM AUTHOR]

Hepatocellular carcinoma is the most common primary malignant hepatic tumor and occurs most often in the setting of chronic liver disease. Liver transplantation is a curative treatment option and is an ideal solution because it solves the chronic underlying liver disorder while removing the malignant lesion. However, due to organ shortages, this treatment can only be applied to carefully selected patients according to clinical guidelines. Artificial intelligence is an emerging technology with multiple applications in medicine with a predilection for domains that work with medical imaging, like radiology. With the help of these technologies, laborious tasks can be automated, and new lesion imaging criteria can be developed based on pixel-level analysis. Our objectives are to review the developing AI applications that could be implemented to better stratify liver transplant candidates. The papers analysed applied AI for liver segmentation, evaluation of steatosis, sarcopenia assessment, lesion detection, segmentation, and characterization. A liver transplant is an optimal treatment for patients with hepatocellular carcinoma in the setting of chronic liver disease. Furthermore, AI could provide solutions for improving the management of liver transplant candidates to improve survival. [ABSTRACT FROM AUTHOR]

Electrical impedance tomography (EIT) is a non-invasive imaging modality used for estimating the conductivity of an object Ω from boundary electrode measurements. In recent years, researchers achieved substantial progress in analytical and numerical methods for the EIT inverse problem. Despite the success, numerical instability is still a major hurdle due to many factors, including the discretization error of the problem. Furthermore, most algorithms with good performance are relatively time consuming and do not allow real-time applications. In our approach, the goal is to separate the unknown conductivity into two regions, namely the region of homogeneous background conductivity and the region of non-homogeneous conductivity. Therefore, we pose and solve the problem of shape reconstruction using machine learning. We propose a novel and simple jet intriguing neural network architecture capable of solving the EIT inverse problem. It addresses previous difficulties, including instability, and is easily adaptable to other ill-posed coefficient inverse problems. That is, the proposed model estimates the probability for a point of whether the conductivity belongs to the background region or to the non-homogeneous region on the continuous space R d ∩ Ω with d ∈ { 2 , 3 } . The proposed model does not make assumptions about the forward model and allows for solving the inverse problem in real time. The proposed machine learning approach for shape reconstruction is also used to improve gradient-based methods for estimating the unknown conductivity. In this paper, we propose a piece-wise constant reconstruction method that is novel in the inverse problem setting but inspired by recent approaches from the 3D vision community. We also extend this method into a novel constrained reconstruction method. We present extensive numerical experiments to show the performance of the architecture and compare the proposed method with previous analytic algorithms, mainly the monotonicity-based shape reconstruction algorithm and iteratively regularized Gauss–Newton method. [ABSTRACT FROM AUTHOR]

In recent years, the rapid development of deep learning approaches has paved the way to explore the underlying factors that explain the data. In particular, several methods have been proposed to learn to identify and disentangle these underlying explanatory factors in order to improve the learning process and model generalization. However, extracting this representation with little or no supervision remains a key challenge in machine learning. In this paper, we provide a theoretical outlook on recent advances in the field of unsupervised representation learning with a focus on auto-encoding-based approaches and on the most well-known supervised disentanglement metrics. We cover the current state-of-the-art methods for learning disentangled representation in an unsupervised manner while pointing out the connection between each method and its added value on disentanglement. Further, we discuss how to quantify disentanglement and present an in-depth analysis of associated metrics. We conclude by carrying out a comparative evaluation of these metrics according to three criteria, (i) modularity, (ii) compactness and (iii) informativeness. Finally, we show that only the Mutual Information Gap score (MIG) meets all three criteria. [ABSTRACT FROM AUTHOR]

Deep learning is a popular topic among machine learning researchers nowadays, with great strides being made in recent years to develop robust artificial neural networks for faster convergence to a reasonable accuracy. Network architecture and hyperparameters of the model are fundamental aspects of model convergence. One such important parameter is the initial values of weights, also known as weight initialization. In this paper, we perform two research tasks concerned with the weights of neural networks. First, we develop three novel weight initialization algorithms inspired by the neuroscientific construction of the mammalian brains and then test them on benchmark datasets against other algorithms to compare and assess their performance. We call these algorithms the lognormal weight initialization, modified lognormal weight initialization, and skewed weight initialization. We observe from our results that these initialization algorithms provide state-of-the-art results on all of the benchmark datasets. Second, we analyze the influence of training an artificial neural network on its weight distribution by measuring the correlation between the quantitative metrics of skewness and kurtosis against the model accuracy using linear regression for different weight initializations. Results indicate a positive correlation between network accuracy and skewness of the weight distribution but no affirmative relation between accuracy and kurtosis. This analysis provides further insight into understanding the inner mechanism of neural network training using the shape of weight distribution. Overall, the works in this paper are the first of their kind in incorporating neuroscientific knowledge into the domain of artificial neural network weights. [ABSTRACT FROM AUTHOR]

Time series classification models have been garnering significant importance in the research community. However, not much research has been done on generating adversarial samples for these models. These adversarial samples can become a security concern. In this paper, we propose utilizing an adversarial transformation network (ATN) on a distilled model to attack various time series classification models. The proposed attack on the classification model utilizes a distilled model as a surrogate that mimics the behavior of the attacked classical time series classification models. Our proposed methodology is applied onto 1-nearest neighbor dynamic time warping (1-NN DTW) and a fully convolutional network (FCN), all of which are trained on 42 University of California Riverside (UCR) datasets. In this paper, we show both models were susceptible to attacks on all 42 datasets. When compared to Fast Gradient Sign Method, the proposed attack generates a larger faction of successful adversarial black-box attacks. A simple defense mechanism is successfully devised to reduce the fraction of successful adversarial samples. Finally, we recommend future researchers that develop time series classification models to incorporating adversarial data samples into their training data sets to improve resilience on adversarial samples. [ABSTRACT FROM AUTHOR]

The digital industrial revolution calls for smart manufacturing plants, i.e. plants that include sensors and vision systems accompanied with artificial intelligence and advanced data analytics in order to meet the required accuracy, reliability and productivity levels. In this paper, we introduce a surface analysis and classification approach based on a deep learning algorithm. The approach is intended to let machining centres recognise the adequacy of process parameters adopted for the milling operation performed, based on the phenomenological effects left on the machined surface. Indeed, the operator will be able to understand how to change process parameters to improve workpiece quality of subsequent parts by a reverse engineering procedure that reconstructs the process parameters that generated the analysed surface. A shallow convolutional neural network was proposed to work on surface image patches based on a limited training dataset of optimal and undesired cutting conditions. The architecture consists of a series of 3 stacked convolutional blocks. The performance of the proposed solution was validated through 5-fold cross-validation, measuring the mean and standard deviation of the f1-score metric. The algorithm arrived at outperformed the best state-of-the-art approach by 4.8% when considering average classification performance. [ABSTRACT FROM AUTHOR]

This paper explores the use of a variational autoencoder to predict the embankment settlement and pore water pressure based on monitoring data. A variational autoencoder can learn intrinsic patterns in embankment behavior through unsupervised deep machine learning. The proposed approach implemented the observational method efficiently because updating soil parameters was no longer a necessary step, unlike in previous research. The embankment response was predicted directly through Gibbs sampling, which involves an iterative encoding and decoding process in the variational autoencoder. The variational autoencoder was trained using simulated embankment responses from the numerical (Plaxis) model. The approach was applied at the Ballina site to predict the embankment response, based on monitoring data with varying time periods. The prediction intervals captured the actual trends satisfactorily, with the intervals becoming more aligned with actual values as more monitoring data were incorporated. The predictions were also more reasonable, compared to those based entirely on representative soil parameters from laboratory or in-situ tests. The variational autoencoder was also applied to another case involving synthetic monitoring data based on the Ballina site, which demonstrates the capability of the variational autoencoder to predict multiple scenarios of embankment behavior. [ABSTRACT FROM AUTHOR]

Quantile regression is a fundamental problem in statistical learning motivated by a need to quantify uncertainty in predictions, or to model a diverse population without being overly reductive. For instance, epidemiological forecasts, cost estimates, and revenue predictions all benefit from being able to quantify the range of possible values accurately. As such, many models have been developed for this problem over many years of research in statistics, machine learning, and related fields. Rather than proposing yet another (new) algorithm for quantile regression we adopt a meta viewpoint: we investigate methods for aggregating any number of conditional quantile models, in order to improve accuracy and robustness. We consider weighted ensembles where weights may vary over not only individual models, but also over quantile levels, and feature values. All of the models we consider in this paper can be fit using modern deep learning toolkits, and hence are widely accessible (from an implementation point of view) and scalable. To improve the accuracy of the predicted quantiles (or equivalently, prediction intervals), we develop tools for ensuring that quantiles remain monotonically ordered, and apply conformal calibration methods. These can be used without any modification of the original library of base models. We also review some basic theory surrounding quantile aggregation and related scoring rules, and contribute a few new results to this literature (for example, the fact that post sorting or post isotonic regression can only improve the weighted interval score). Finally, we provide an extensive suite of empirical comparisons across 34 data sets from two different benchmark repositories. [ABSTRACT FROM AUTHOR]

Machine learning algorithms are widely used in many intelligent applications and cloud services. Currently, the hottest topic in this field is Deep Learning represented often by neural network structures. Deep learning is fully known as deep neural network, and artificial neural network is a typical machine learning method and an important way of deep learning. With the massive growth of data, deep learning research has made significant achievements and is widely used in natural language processing (NLP), image recognition, and autonomous driving. However, there are still many breakthroughs needed in the training time and energy consumption of deep learning. Based on our previous research on fast learning architecture for neural network, in this paper, a solution to minimize the learning time of a fully connected neural network is analysed theoretically. Therefore, we propose a new parallel algorithm structure and a training method with over-tuned parameters. This strategy finally leads to an adaptation delay and the impact of this delay on the learning performance is analyzed using a simple benchmark case study. It is shown that a reduction of the adaptation step size could be proposed to compensate errors due to the delayed adaptation, then the gain in processing time for the learning phase is analysed as a function of the network parameters chosen in this study. Finally, to realize the real-time learning, this solution is implemented with a FPGA due to the parallelism architecture and flexibility, this integration shows a good performance and low power consumption. [ABSTRACT FROM AUTHOR]

This article focuses on the problem of detecting disinformation about COVID-19 in online discussions. As the Internet expands, so does the amount of content on it. In addition to content based on facts, a large amount of content is being manipulated, which negatively affects the whole society. This effect is currently compounded by the ongoing COVID-19 pandemic, which caused people to spend even more time online and to get more invested in this fake content. This work brings a brief overview of how toxic information looks like, how it is spread, and how to potentially prevent its dissemination by early recognition of disinformation using deep learning. We investigated the overall suitability of deep learning in solving problem of detection of disinformation in conversational content. We also provided a comparison of architecture based on convolutional and recurrent principles. We have trained three detection models based on three architectures using CNN (convolutional neural networks), LSTM (long short-term memory), and their combination. We have achieved the best results using LSTM (F1 = 0.8741, Accuracy = 0.8628). But the results of all three architectures were comparable, for example the CNN+LSTM architecture achieved F1 = 0.8672 and Accuracy = 0.852. The paper offers finding that introducing a convolutional component does not bring significant improvement. In comparison with our previous works, we noted that from all forms of antisocial posts, disinformation is the most difficult to recognize, since disinformation has no unique language, such as hate speech, toxic posts etc. [ABSTRACT FROM AUTHOR]

It is widely acknowledged that the success of deep learning is built upon large-scale training data and tremendous computing power. However, the data and computing power are not always available for many real-world applications. In this paper, we address the machine learning problem where it lacks training data and limits computing power. Specifically, we investigate domain adaptation which is able to transfer knowledge from one labeled source domain to an unlabeled target domain, so that we do not need much training data from the target domain. At the same time, we consider the situation that the running environment is confined, e.g., in edge computing the end device has very limited running resources. Technically, we present the Faster Domain Adaptation (FDA) protocol and further report two paradigms of FDA: early stopping and amid skipping. The former accelerates domain adaptation by multiple early exit points. The latter speeds up the adaptation by wisely skip several amid neural network blocks. Extensive experiments on standard benchmarks verify that our method is able to achieve the comparable and even better accuracy but employ much less computing resources. To the best of our knowledge, there are very few works which investigated accelerating knowledge adaptation in the community. This work is expected to inspire the topic for more discussion. [ABSTRACT FROM AUTHOR]

Classification of 3D objects – the selection of a category in which each object belongs – is of great interest in the field of machine learning. Numerous researchers use deep neural networks to address this problem, altering the network architecture and representation of the 3D shape used as an input. To investigate the effectiveness of their approaches, we conduct an extensive survey of existing methods and identify common ideas by which we categorize them into a taxonomy. Second, we evaluate 11 selected classification networks on two 3D object datasets, extending the evaluation to a larger dataset on which most of the selected approaches have not been tested yet. For this, we provide a framework for converting shapes from common 3D mesh formats into formats native to each network, and for training and evaluating different classification approaches on this data. Despite being partially unable to reach the accuracies reported in the original papers, we compare the relative performance of the approaches as well as their performance when changing datasets as the only variable to provide valuable insights into performance on different kinds of data. We make our code available to simplify running training experiments with multiple neural networks with different prerequisites. [ABSTRACT FROM AUTHOR]

Active learning is focused on minimizing the effort required to obtain labeled data by iteratively choosing fresh data samples for training a machine learning model. One of the primary challenges in active learning involves the selection of the most informative instances for labeling by an annotation oracle at each iteration. A viable approach is to develop an active learning strategy that aligns with the performance of a meta-learning model. This strategy evaluates the quality of previously selected instances and subsequently trains a machine learning model to predict the quality of instances to be labeled in the current iteration. This paper introduces a novel approach to learning for active learning, wherein instances are chosen for labeling based on their potential to induce the most substantial change in the current classifier. We explore various strategies for assessing the significance of an instance, taking into account variations in the learning gradient of the classification model. Our approach can be applied to any classifier that can be trained using gradient descent optimization. Here, we present a formulation that leverages a deep neural network model, which has not been extensively explored in existing learning-to-active-learn methodologies. Through experimental validation, our approach demonstrates promising results, especially in scenarios where there are limited initially labeled instances, particularly when the number of labeled instances per class is extremely limited. • A new meta-active learning strategy for deep neural network models. • The instance selection problem is modeled as a regression problem. • Four strategies of instance importance scoring by considering gradient variations. • Experiments have shown good results when few initially labeled objects are available. • A dimensionality reduction step for the instances improves the performances. [ABSTRACT FROM AUTHOR]

Sparsity-constrained optimization problems are common in machine learning, such as sparse coding, low-rank minimization and compressive sensing. However, most of previous studies focused on constructing various hand-crafted sparse regularizers, while little work was devoted to learning adaptive sparse regularizers from given input data for specific tasks. In this paper, we propose a deep sparse regularizer learning model that learns data-driven sparse regularizers adaptively. Via the proximal gradient algorithm, we find that the sparse regularizer learning is equivalent to learning a parameterized activation function. This encourages us to learn sparse regularizers in the deep learning framework. Therefore, we build a neural network composed of multiple blocks, each being differentiable and reusable. All blocks contain learnable piecewise linear activation functions which correspond to the sparse regularizer to be learned. Furthermore, the proposed model is trained with back propagation, and all parameters in this model are learned end-to-end. We apply our framework to multi-view clustering and semi-supervised classification tasks to learn a latent compact representation. Experimental results demonstrate the superiority of the proposed framework over state-of-the-art multi-view learning models. [ABSTRACT FROM AUTHOR]

The efficiency of link adaptation in wireless communications relies greatly on the accuracy of channel knowledge and transmission mode selection. In this paper, a novel deep learning based link adaptation framework is proposed for the orthogonal frequency-division multiplexing (OFDM) systems with compressed-sensing-assisted index modulation, termed as OFDM-CSIM, communicating over millimeter-wave (mmWave) channels. To achieve link adaptation, a novel multi-layer sparse Bayesian learning (SBL) algorithm is proposed for accurately and instantaneously providing the required channel state information. Meanwhile, a deep neural networks (DNN)-assisted adaptive modulation algorithm is proposed to choose the best possible transmission mode to maximize the achievable throughput. Simulation results show that the proposed multi-layer SBL algorithm enables more accurate channel estimation than the conventional techniques. The DNN-based adaptive modulator is capable of achieving a higher throughput than the learning-assisted solution based on the $k$ nearest neighbor ($k$ -NN) algorithm, and also the classic average signal-to-noise ratio (SNR)-based solutions. Moreover, analysis shows that both the multi-layer SBL algorithm and the DNN-assisted adaptive modulator achieve better performance than their respective conventional counterparts while at a significantly lower computational complexity cost. [ABSTRACT FROM AUTHOR]

Deep convolutional neural networks (DCNNs) are currently the method of choice both for generative, as well as for discriminative learning in computer vision and machine learning. The success of DCNNs can be attributed to the careful selection of their building blocks (e.g., residual blocks, rectifiers, sophisticated normalization schemes, to mention but a few). In this paper, we propose $\Pi$ Π -Nets, a new class of function approximators based on polynomial expansions. $\Pi$ Π -Nets are polynomial neural networks, i.e., the output is a high-order polynomial of the input. The unknown parameters, which are naturally represented by high-order tensors, are estimated through a collective tensor factorization with factors sharing. We introduce three tensor decompositions that significantly reduce the number of parameters and show how they can be efficiently implemented by hierarchical neural networks. We empirically demonstrate that $\Pi$ Π -Nets are very expressive and they even produce good results without the use of non-linear activation functions in a large battery of tasks and signals, i.e., images, graphs, and audio. When used in conjunction with activation functions, $\Pi$ Π -Nets produce state-of-the-art results in three challenging tasks, i.e., image generation, face verification and 3D mesh representation learning. The source code is available at https://github.com/grigorisg9gr/polynomial_nets. [ABSTRACT FROM AUTHOR]

With the popularity of location-aware devices (e.g., smart phones), large amounts of location-based social media data such as check-ins are generated. This stimulates plenty of studies for POI recommendations by applying machine learning techniques. However, most of the existing studies focus on POI recommendations in the same city or region, and fail to recommend POIs for users when they travel to a new city. In this paper, we propose a novel deep neural network, named as ST-TransRec, for crossing-city POI recommendations. It integrates the deep neural network, transfer learning technique, and density-based resampling method into a unified framework. In this model, the deep neural network is used to capture users’ preferences for POIs and learn the embeddings of POIs. Besides, the transfer learning technique is employed to bridge the gap between cities that results from the city-dependent features. As the distributions over POIs are imbalanced, we design a density-based spatial resampling model which enables POIs to be well matched across cities. We conduct extensive experiments on two real-world datasets. The experimental results show the advantages of ST-TransRec over the state-of-the-art methods for crossing-city POI recommendations. [ABSTRACT FROM AUTHOR]

A self-organization hydrodynamic process has recently been proposed to partially explain the formation of femtosecond laser-induced nanopatterns on Nickel, which have important applications in optics, microbiology, medicine, etc. Exploring laser pattern space is difficult, however, which simultaneously (i) motivates using machine learning (ML) to search for novel patterns and (ii) hinders it, because of the few data available from costly and time-consuming experiments. In this paper, we use ML to predict novel patterns by integrating partial physical knowledge in the form of the Swift-Hohenberg (SH) partial differential equation (PDE). To do so, we propose a framework to learn with few data, in the absence of initial conditions, by benefiting from background knowledge in the form of a PDE solver. We show that in the case of a self-organization process, a feature mapping exists in which initial conditions can safely be ignored and patterns can be described in terms of PDE parameters alone, which drastically simplifies the problem. In order to apply this framework, we develop a second-order pseudospectral solver of the SH equation which offers a good compromise between accuracy and speed. Our method allows us to predict new nanopatterns in good agreement with experimental data. Moreover, we show that pattern features are related, which imposes constraints on novel pattern design, and suggest an efficient procedure of acquiring experimental data iteratively to improve the generalization of the learned model. It also allows us to identify the limitations of the SH equation as a partial model and suggests an improvement to the physical model itself. [ABSTRACT FROM AUTHOR]

The proliferation of machine learning services in the last few years has raised data privacy concerns. Homomorphic encryption (HE) enables inference using encrypted data but it incurs 100x–10,000x memory and runtime overheads. Secure deep neural network (DNN) inference using HE is currently limited by computing and memory resources, with frameworks requiring hundreds of gigabytes of DRAM to evaluate small models. To overcome these limitations, in this paper we explore the feasibility of leveraging hybrid memory systems comprised of DRAM and persistent memory. In particular, we explore the recently-released Intel® Optane™ PMem technology and the Intel® HE-Transformer nGraph® to run large neural networks such as MobileNetV2 (in its largest variant) and ResNet-50 for the first time in the literature. We present an in-depth analysis of the efficiency of the executions with different hardware and software configurations. Our results conclude that DNN inference using HE incurs on friendly access patterns for this memory configuration, yielding efficient executions. [ABSTRACT FROM AUTHOR]

The detection of degraded soil distribution areas is an urgent task. It is difficult and very time consuming to solve this problem using ground methods. The modeling of degradation processes based on digital elevation models makes it possible to construct maps of potential degradation, which may differ from the actual spatial distribution of degradation. The use of remote sensing data (RSD) for soil degradation detection is very widespread. Most often, vegetation indices (indicative botany) have been used for this purpose. In this paper, we propose a method for constructing soil maps based on a multi-temporal analysis of the bare soil surface (BSS). It is an alternative method to the use of vegetation indices. The detection of the bare soil surface was carried out using the spectral neighborhood of the soil line (SNSL) technology. For the automatic recognition of BSS on each RSD image, computer vision based on deep machine learning (neural networks) was used. A dataset of 244 BSS distribution masks on 244 Landsat 4, 5, 7, and 8 scenes over 37 years was developed. Half of the dataset was used as a training sample (Landsat path/row 173/028). The other half was used as a test sample (Landsat path/row 174/027). Binary masks were sufficient for recognition. For each RSD pixel, value "1" was set when determining the BSS. In the absence of BSS, value "0" was set. The accuracy of the machine prediction of the presence of BSS was 75%. The detection of degradation was based on the average long-term spectral characteristics of the RED and NIR bands. The coefficient Cmean, which is the distance of the point with the average long-term values of RED and NIR from the origin of the spectral plane RED/NIR, was calculated as an integral characteristic of the mean long-term values. Higher long-term average values of spectral brightness served as indicators of the spread of soil degradation. To test the method of constructing soil degradation maps based on deep machine learning, an acceptance sample of 133 Landsat scenes of path/row 173/026 was used. On the territory of the acceptance sample, ground verifications of the maps of the coefficient Cmean were carried out. Ground verification showed that the values of this coefficient make it possible to estimate the content of organic matter in the plow horizon (R2 = 0.841) and the thickness of the humus horizon (R2 = 0.8599). In total, 80 soil pits were analyzed on an area of 649 ha on eight agricultural fields. Type I error (false positive) of degradation detection was 17.5%, and type II error (false negative) was 2.5%. During the determination of the presence of degradation by ground methods, 90% of the ground data coincided with the detection of degradation from RSD. Thus, the quality of machine learning for BSS recognition is sufficient for the construction of soil degradation maps. The SNSL technology allows us to create maps of soil degradation based on the long-term average spectral characteristics of the BSS. [ABSTRACT FROM AUTHOR]

Power quality disturbances (PQD) have a negative impact on power quality-sensitive equipment, often resulting in great financial losses. To prevent these losses, besides detecting a PQD on time, it is important to classify it, so that appropriate recovery procedures are employed. The majority of research employs machine learning model PQD classifiers on manually extracted features from simulated or real-world signals. This paper presents an end-to-end approach that circumvents the manual feature extraction and uses signals generated from mathematical voltage sag type formulas. We developed a configurable voltage sag generator that was used to form training and validation datasets. Based on the synthetic three-phase voltage signals, we trained several end-to-end LSTM classifiers that classify voltage sags according to ABC classification. The best-performing model achieved an accuracy of over 90% in the real-world dataset. [ABSTRACT FROM AUTHOR]

Fully connected (FC) layers are used in almost all neural network architectures ranging from multilayer perceptrons to deep neural networks. FC layers allow any kind of symmetric/asymmetric interaction between features without making any assumption about the structure of the data. However, success of convolutional and recursive layers and findings of many studies have proven that the intrinsic structure of a dataset holds a great potential to improve the success of a classification problem. Leveraging clustering to explore and exploit this intrinsic structure in classification problems has been the subject of various studies. In this paper, we propose a new training pipeline for fully connected layers which enables them to make more accurate classification predictions. The proposed method aims to reflect the clustering patterns in the original feature space of the training dataset to the transformed feature space created by the FC layer. In this way, we intend to enhance the representation ability of the extracted features and accordingly increase the classification accuracy. The Fuzzy C-Means algorithm is employed in this study as the clustering tool. To evaluate the performance of the proposed method, 11 experiments were conducted on 9 benchmark UCI datasets. Empirical results show that the proposed method works well in practice and gives higher classification accuracies compared to a regular FC layer in most datasets. [ABSTRACT FROM AUTHOR]

Many learning algorithms such as kernel machines, nearest neighbors, clustering, or anomaly detection, are based on distances or similarities. Before similarities are used for training an actual machine learning model, we would like to verify that they are bound to meaningful patterns in the data. In this paper, we propose to make similarities interpretable by augmenting them with an explanation. We develop BiLRP, a scalable and theoretically founded method to systematically decompose the output of an already trained deep similarity model on pairs of input features. Our method can be expressed as a composition of LRP explanations, which were shown in previous works to scale to highly nonlinear models. Through an extensive set of experiments, we demonstrate that BiLRP robustly explains complex similarity models, e.g., built on VGG-16 deep neural network features. Additionally, we apply our method to an open problem in digital humanities: detailed assessment of similarity between historical documents, such as astronomical tables. Here again, BiLRP provides insight and brings verifiability into a highly engineered and problem-specific similarity model. [ABSTRACT FROM AUTHOR]

Parkinson's disease (PD) is a neurodegenerative disorder that produces both motor and non-motor complications, degrading the quality of life of PD patients. Over the past two decades, the use of wearable devices in combination with machine learning algorithms has provided promising methods for more objective and continuous monitoring of PD. Recent advances in artificial intelligence have provided new methods and algorithms for data analysis, such as deep learning (DL). The aim of this article is to provide a comprehensive review of current applications where DL algorithms are employed for the assessment of motor and non-motor manifestations (NMM) using data collected via wearable sensors. This paper provides the reader with a summary of the current applications of DL and wearable devices for the diagnosis, prognosis, and monitoring of PD, in the hope of improving the adoption, applicability, and impact of both technologies as support tools. Following PRISMA (Systematic Reviews and Meta-Analyses) guidelines, sixty-nine studies were selected and analyzed. For each study, information on sample size, sensor configuration, DL approaches, validation methods and results according to the specific symptom under study were extracted and summarized. Furthermore, quality assessment was conducted according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) method. The majority of studies (74%) were published within the last three years, demonstrating the increasing focus on wearable technology and DL approaches for PD assessment. However, most papers focused on monitoring (59%) and computer-assisted diagnosis (37%), while few papers attempted to predict treatment response. Motor symptoms (86%) were treated much more frequently than NMM (14%). Inertial sensors were the most commonly used technology, followed by force sensors and microphones. Finally, convolutional neural networks (52%) were preferred to other DL approaches, while extracted features (38%) and raw data (37%) were similarly used as input for DL models. The results of this review highlight several challenges related to the use of wearable technology and DL methods in the assessment of PD, despite the advantages this technology could bring in the development and implementation of automated systems for PD assessment. • Deep learning and wearable technology are used in diagnosis and monitoring. • Growing interest in the use of deep learning and wearables in Parkinson's disease. • The adoption of deep learning in Parkinson's disease is currently under development. • Technical and methodological challenges difficult the technology adoption. [ABSTRACT FROM AUTHOR]

Seizure prediction for drug-refractory epilepsy patients can improve their quality of life, reduce their anxiety, and help them take the necessary precautions. Nowadays, numerous deep learning algorithms have been proposed to predict seizure onset and obtain better performance than traditional machine learning methods. However, these methods require a large set of parameters and large hardware resources; they also have high energy consumption. Therefore, these methods cannot be implemented on compact, low-power wearable, or implantable medical devices. The devices should operate on a real-time basis to continually inform the epileptic patients. In this paper, we describe energy-efficient and hardware-friendly methods to predict the epileptic seizures. A model of only 45 kB was obtained by the neural architecture search and was evaluated across three datasets. The overall sensitivity, false prediction rate, and area under receiver operating characteristic curve were 99.81%, 0.005/h, 1 and 93.48%, 0.063/h, 0.977 and 85.19%, 0.116/h, 0.933, respectively, for three public datasets. This model was further reduced with several methods for model compression. The performances for the model sizes less than 50 kB for scalp EEG data and less than 10 kB for intracranial EEG data outperformed all the other models of similar model sizes. In particular, the energy consumption estimation was less than 10 mJ per inference for scalp EEG signal and less than 0.5 mJ per inference for intracranial EEG signal, which meet the requirements for low-power wearable and implantable devices, respectively. [ABSTRACT FROM AUTHOR]

Multi-view representation learning is a promising and challenging research topic, which aims to integrate multiple data information from different views to improve the learning performance. The recent deep Gaussian processes (DGPs) have the advantages of good uncertainty estimates, powerful non-linear mapping ability and great generalization capability, which can be used as an excellent data representation learning method. However, DGPs only focus on single view data and are rarely applied to the multi-view scenario. In this paper, we propose a multi-view representation learning algorithm with deep Gaussian processes (named MvDGPs), which inherits the advantages of deep Gaussian processes and multi-view representation learning, and can learn more effective representation of multi-view data. The MvDGPs consist of two stages. The first stage is multi-view data representation learning, which is mainly used to learn more comprehensive representations of multi-view data. The second stage is classifier design, which aims to select an appropriate classifier to better employ the representations obtained in the first stage. In contrast with DGPs, MvDGPs support asymmetrical modeling depths for different views of data, resulting in better characterizations of the discrepancies among different views. Experimental results on real-world multi-view data sets verify the effectiveness of the proposed algorithm, which indicates that MvDGPs can integrate the complementary information in multiple views to discover a good representation of the data. [ABSTRACT FROM AUTHOR]

This paper describes the development of U-net++ models, a type of neural network that performs deep learning, to emulate the shortwave Rapid Radiative Transfer Model (RRTM). The goal is to emulate the RRTM accurately in a small fraction of the computing time, creating a U-net++ that could be used as a parameterization in numerical weather prediction (NWP). Target variables are surface downwelling flux, top-of-atmosphere upwelling flux ( F up TOA ), net flux, and a profile of radiative-heating rates. We have devised several ways to make the U-net++ models knowledge-guided, recently identified as a key priority in machine learning (ML) applications to the geosciences. We conduct two experiments to find the best U-net++ configurations. In experiment 1, we train on nontropical sites and test on tropical sites, to assess extreme spatial generalization. In experiment 2, we train on sites from all regions and test on different sites from all regions, with the goal of creating the best possible model for use in NWP. The selected model from experiment 1 shows impressive skill on the tropical testing sites, except four notable deficiencies: large bias and error for heating rate in the upper stratosphere, unreliable F up TOA for profiles with single-layer liquid cloud, large heating-rate bias in the midtroposphere for profiles with multilayer liquid cloud, and negative bias at low zenith angles for all flux components and tropospheric heating rates. The selected model from experiment 2 corrects all but the first deficiency, and both models run ~104 times faster than the RRTM. Our code is available publicly. [ABSTRACT FROM AUTHOR]

In this paper, we investigate the detection problem for a massive machine-type communication (mMTC) system that has correlated user activities. Two deep learning assisted algorithms are proposed to exploit the user activity correlation to facilitate channel estimation and user identification. Due to the dependency among user activities, conventional element-wise minimum mean square error (MMSE) denoiser used in the orthogonal approximate message passing (OAMP) algorithm cannot achieve satisfying performance during the two-step iterative process. Therefore, we propose a deep learning modified OAMP (DL-mOAMP) algorithm, which iteratively modifies the user activity ratio via exploiting the user activity correlation in the MMSE denoiser based on the estimated sequence during each OAMP iteration. Moreover, given a specific false alarm probability, a constant threshold employed in the conventional user identification is not optimal in the presence of user activity correlation. Thus, we propose a neural network framework that is dedicated to the user identification (DL-mOAMP-UI algorithm), which minimizes the missed detection probability under a pre-determined false alarm probability. Numerical results show that the proposed DL-mOAMP algorithm provides a substantial mean squared error performance gain compared to the conventional OAMP algorithm and the DL-mOAMP-UI algorithm can further improve the user identification accuracy of an mMTC system. [ABSTRACT FROM AUTHOR]

Background: Detection of pulmonary nodules is an important aspect of an automatic detection system. Incomputer-aided diagnosis (CAD) systems, the ability to detect pulmonary nodules is highly important, which plays an important role in the diagnosis and early treatment of lung cancer. Currently, the detection of pulmonary nodules depends mainly on doctor experience, which varies. This paper aims to address the challenge of pulmonary nodule detection more effectively. Methods: A method for detecting pulmonary nodules based on an improved neural network is presented in this paper. Nodules are clusters of tissue with a diameter of 3 mm to 30 mm in the pulmonary parenchyma. Because pulmonary nodules are similar to other lung structures and have a low density, false positive nodules often occur. Thus, our team proposed an improved convolutional neural network (CNN) framework to detect nodules. First, a nonsharpening mask is used to enhance the nodules in computed tomography (CT) images; then, CT images of 512×512 pixels are segmented into smaller images of 96×96 pixels. Second, in the 96×96 pixel images which contain or exclude pulmonary nodules, the plaques corresponding to positive and negative samples are segmented. Third, CT images segmented into 96×96 pixels are down-sampled to 64×64 and 32×32 size respectively. Fourth, an improved fusion neural network structure is constructed that consists of three three-dimensional convolutional neural networks, designated as CNN-1, CNN-2, and CNN-3, to detect false positive pulmonary nodules. The networks’ input sizes are 32×32×32, 64×64×64, and 96×96×96 and include 5, 7, and 9 layers, respectively. Finally, we use the AdaBoost classifier to fuse the results of CNN-1, CNN-2, and CNN-3. We call this new neural network framework the Amalgamated-Convolutional Neural Network (A-CNN) and use it to detect pulmonary nodules. Findings: Our team trained A-CNN using the LUNA16 and Ali Tianchi datasets and evaluated its performance using the LUNA16 dataset. We discarded nodules less than 5mm in diameter. When the average number of false positives per scan was 0.125 and 0.25, the sensitivity of A-CNN reached as high as 81.7% and 85.1%, respectively. [ABSTRACT FROM AUTHOR]